Lehrstuhl für Bildverarbeitung

Institute of Imaging & Computer Vision

Noise-Resistant Weak-StructureEnhancement for Digital Radiography

Martin Stahl and Til Aach and Sabine Dippel and Thorsten Buzugand Raphael Wiemker and Ulrich Neitzel

in: SPIE Vol. 3661: Medical Imaging 99: Image Processing. See also BIBTEX entry below.

BIBTEX:@inproceedings{STA99a,author = {Martin Stahl and Til Aach and Sabine Dippel

and Thorsten Buzug and Raphael Wiemker andUlrich Neitzel},

title = {Noise-Resistant Weak-Structure Enhancement forDigital Radiography},

booktitle = {{SPIE Vol. 3661: Medical Imaging 99: Image Processing}},editor = {K. M. Hanson},publisher = {SPIE},address = {San Diego, USA},month = {February 20--26},year = {1999},pages = {1406--1417}}

© 1999 Society of Photo-Optical Instrumentation Engineers. This paper was published in SPIEVol. 3661: Medical Imaging 99: Image Processing and is made available as an electronic reprintwith permission of SPIE. One print or electronic copy may be made for personal use only.Systematic or multiple reproduction, distribution to multiple locations via electronic or othermeans, duplication of any material in this paper for a fee or for commercial purposes, ormodification of the content of the paper are prohibited.

document created on: December 19, 2006created from file: spiemi99coverpage.texcover page automatically created with CoverPage.sty(available at your favourite CTAN mirror)

1\

PROCEEDINCS OF SP IEwSPIE-The Internat ional Society for Opt ica l Engineer ing

Medical Imaging 1999

Image Processi ng

K e n n e t h M . H a n s o nCha i r lEd i to r

22-25 February 1999San D iego, Ca l i fo rn ia

Sponso red byS P I E - T h e I n t e r n a t i o n a l S o c i e t y f o r O p t i c a l E n g i n e e r i n g

Coope r a t i n g O r gan i za t i on sAAPM-Amer i can Assoc ia t i on o f Phys i c i s t s i n Med ic ineAPS-Amer i can Phys io log i ca l Soc ie t yFDA Cen te r f o r Dev i ces and Rad io log i ca l Hea l t hlS&T-The Soc ie t y f o r lmag ing Sc ience and Techno logyN EMA-Na t i ona l E lec t r i ca l Manu fac tu re rs Assoc ia t i on /D ragnos t i c

S y s t e m s D i v i s i o nR S N A - R a d i o l o g i c a l S o c i e t y o f N o r t h A m e r i c aS C A R - S o c i e t y f o r C o m p u t e r A p p l i c a t i o n s i n R a d i o l o g y

Pub l i shed byS P I E - T h e I n t e r n a t i o n a l S o c i e t y f o r O p t i c a l F n g i n e e r i n g

SPIE is an in te rna t iona l techn ica l soc ie ty ded ica ted to advanc ing eng ineer ing and sc ien t i f i capp l i ca t ions o f op t ica l , photon ic , imag ing , e lec t ron ic , and op toe lec t ron ic iechno log ies .

l m a g i n g a n d T h e r a p y

V o l u m e 3 6 6 1Part Two of Two Parts

f l[T PosrenlYllAlranoHt""*eb M^rlt"

Noise-resistant weak-structure enhancement for digitalradiograPhY

N{. Stahl ' . T. AachÖ, T. lVI. Buzugo, S. Dippelo and U. Neitzel ' '

o Phiiips Research, R.oentgenstrasse 24-26, D-22335 Hamburg, Germanv

b Nledical Universi ty of Lübeck. D-23569 Lübeck, GermartY

N4edical Svstems, Roentgenstrasse 24 D-22335 Hamburg, Germany

ABSTRACT

T'da.v's <ligital radiography systems mostly use unsharp masking-like image enhancement techniques based on split-

ting input images into two or three frequency channeis. This rnethod allows to enhance very small structures (edge

.nhun.",n"nt) as well as enhancement of global contrast (harmonization). However, structures of medium size are

not accessible by such enhancement. We develop and test a nonlinear enhancement algorithm based on hierarchi-

cally repeated unsharp m:xking, resulting in a multiscale architecture allowing consistent access to structures of

all sizes. Thc algorithm is noise-resistant in the sense that it prevents unacceptable noise amplification. Clinical

tests performecl in the raciiology departments of two niajor German hospitals so far strongly indicate the superior

performance and high acceptance of the new processittg'

Keywords: digitül radiography, image enhancement, rnulti-resolution techniques, multiscale enhancement, noise

resrstance:

1. INTRODUCTION

In X-radiogr.aphs often large differences in X-ray density dominate over the diagnostically relevant detail information.

Hard- or sofic6p1. display then fuces the clilemma of having to reproduce subtle details with sufficient contrast and

at the same time large variations witirout ciipping. In ciigital radiography, these conflicting requirements can be

reco'cilecl by suitable processing.l 6 The large densitl' variations occur predominantly over low frequencies and

rliagnosticalll, relevant information is mainly high frequent. Therefore, most present systems separate the image tnto

tw..(or t i rree) frequency bancls al lowing selective enhancement of the high frequent information. Receiver operating

charar:terist ir .(ROC) studies show that detai l perception can indeecl be improved by appropriate processing'2

Whiie these irnage enhancement nrethods have been tuned towards considerable performance over the years'

i'lrroved i'rage ciecgrnposition techniques, like waveletsT and pyramidss have emerged. These mzlriscole techniques

rlecom'.sr: irrages into several frequency channels of varying spatial resolution. The advantage of these decomposition

tcchrriq'es rs that structures of different size appear separately in different scales, and can be processed independently'

a.s cl.ne for t . ler, ision images in.e However, when applied to (digital) radiographs (see e.g.10 12), i t turns out that

t6ese appr.acles ar.e as sensitive to noise as unsharp mixking-based algorithms. The reason for this is that the finest

sr:ale is , ' ror(,or less equivalent to the high-frequency channei of the described standard processing, where noise is

r r ros t c r i t i ca l .

Irr tSis paper we therefore focus on a rroise-resistant mult iscale enhancement algori thm- we show that mult iscale

'r.ccssi'g r:arr bc regar6ed a^s hierarchically repeated unsharp masking, and describe how existing unsharp masking-

lrased algoritlrrrrs czrl be rrrapped on a multiscale structure. We then exploit the additional opportunities of this

architerr:ture for coltrast enhancement, arrd describe horv to prevent uncontrolled noise amplification. Finally, results

, , I t r f i r s t c l i r r i ca l s l .u ( lv a r ( , p rese t rLe t l .

F\rr ther : rut l tor informat ion: (Send correspondence to M Stahl)-1. . \ . i . I luzuq: present ar ldress: RheinAhrCampus Remagen, UniVersi ty of Appl ied Sciences, Sueclal le 2,53424 Remagen, (Jerrnany

Pa r t o f t he SP IE Con fe rence on lmaqe P rocess ino o San D ieqo Ca l i f o rn i a r Feb rua rv 1999

SPIE Vo l . 3661 . 0277 -7 86X /99 /$10 .001 406

&

/ . . 2 . FROM UNSHARP MASKING TO MULTISCALE PROCESSING

As i refe."n.e stanilard algorithm to be nrapped later orr a rnultiscale pyrarnid, we use a two-step unsharp masking-

iu".a "fgo.i thnr (Figure 1), which is appl icable e.g. to storage phosphor radiographst: First, a si ightly blurred

version /bl,..e.t, generated by filtering witli a small kernel, is subtracted from the original 1o.*. The resuiting high-

f*qu.n.y image 1"6*" contains mostly edge information, and is enhanced by a constanL sharpness /aclor SF'".

eaäing ihe enhanced highpass data to /br ' , ." , i provides an image which exhibits increased perceived sharpness (-Edge

Enhoncemerrt). Then, a large box kernel - e.g. 201 x 201 pixel for a2k x 2,k-radiograph - is appl ied to separate

middle an6 high spatial frequencies (contrast information, I .o.t ,*t) from the very low ones. Subjecting 1.o.r.o., to

another corrstant factor - termed contrast factor CF,'" - amplifies contrast information relative to the large and

low-frequent clensity r.ariat ions ldensity (Harmonizatlon). Note that nonl inear ampli f icat ion l ikea is easi l f int,roduced

into this structure by choosing density- or contrast-dependent gain factors (see dotted l ines in Figure 1). This holds

..g. io, tire reference stanclarcl algorithm that is used in 3.3 for chest images.13

Figure 1. Two step unsharp-masking-based processing

Obviously, this algorithm aliows separate access only to fine detail and low-frequent information. IUedium-sized

structures remain grouped together in the broad frequency band 1.o.1r""1. The algorithm's performance can hence be

considerably improved b1' accessing structures of arbitrary size in the same way as fine structures. For convenience,

we seek to hierarchically' continue the Edge Enhancement operation, where, to avoid an unpractical increase in the

amount of data, each lowpass image .Ib1u,,g,1 is subsampled by a factor of two in each direction ({). Consequently for

subtraction fronr the original, as well as addition in the reconstruction path, an expansion step must be introduced,

i.e. zeros are inserted between the available samples (f) followed by another lowpass.

Ideally, the lowpass preceding the subsampling should cut off at half the Nyquist frequency. For practical reasons,

only smali binomial filter kertrels are used, thus introducing some aliasing into the subsampled images. However, this

is not critical, because ttre alias cancels out during reconstruction. This balance between produced and cancelled

al ias is disturbed by' subband processing but general iy not to a severe degree.

The subsampiing provides an image 1o,g,51 of half height and half rvidth compared to the original, representing the

spatial frequencies up to half the Nyquist frequency of the original image. The image itself, however, contains spatial

frequencies up to the original Nyquist frequency, because subsampling has doubled all actually present frequencies.

This means that the subsampled image is a sharp but smaller image, where edges correspond to larger structures in

the originai. Therefore, the sane filter as already used in the upper part of Figure 1 can be applied again to process

the edges of 1o,r,.ei. Repeating this nesting of unsharp-masking operations n times leads to the algorithm in Figure

2, which is known as Lht: Laplacian PyrarniiL 8 In the following, the difierence images will be referred to as subband

images or subbands. To ac|ieve backward compatibility of this algorithm to our reference processing, we now map

the two-step linear filtering operation on the pyramid. If 3 x 3 binonria] filter kernels are used at each level of the

pyramid, the finest. uppernlost level of the pyrarnid 56 is almost equivalent to the edge image 1"6gs of Figure 1.

The number of pyrarnid levels needed is deterrnined by the requirement that the (expanded) iowpass level generated

1407

I ' o t s

r - _ - - r" . r r l l l . l l r l l - - ; " L - ILtr;;F-

'"o*lF- s o

I s i n ( . V L r tr r \ u /

- I ^ , I

. \ s l l l ( 7 I . i t J

sF-

Figure 2. The Laplacian Pyramid with subband processing modules

0.002

Figure 3. Conrparison of the N'ITF of a one dimensional box kernel I I(u) to a pyramid transfer function G(u)

considering onlv t ire lo\vpass level. Tire length of the box kerrrel is Ä : 201 pixeis. Tl ie pyramid consists of 8 levels,

each usinp5 a birromial kernel with a length of 3 pixels.

within tire pvramid is approximately equivalent to I,1g,,511, in Figure 1 . This is met if the magnitr.rde transfer function(N,{TF) of t}re pyramid, considering only the lowpass level, is a goocl approximation of the N{TF of the large box

kernel. The N{TF H(u') of a box kernel f i l ter of size N x N is given b"'

0.8

0.6

o.2

( l )

- l lI org,s ln-2) 1 |

| | en.S{n_2) |

T r - - 1WA-@r-ä*@l-E

The lowpass filter chosen for the pyramid is a separable binomial filter of size 3 x 3 pixeis, basecl c,n the one dimensionai

kerne l [0 .25 0 .5 0 .25 ] . I t s IUTF G; (u) i s g iven by

G y Q r ) : ( c o s ( r T i ) ) 2 ( 2 )

an n-level pvrarnicl arrd including tlie effects of spectral spreading caused byApplving this recursivelr

subsampling yields the transfer function for the lowpass level of the pvramid. For a comparison rvith the transferfunction of the box kernel, it must be taken into account that the reconstruction path of tire pyramid applies thesame sequence of filters for interpolation. Therefore, the trartsfer function for the lowpa-ss level of the pyrarnid mustbe squared to obtain the transfer function for the expanded lowpass levei:

/ r t - ) \ '

G(u) = { T f c r (z ' u ) I\ ä / '

( 3 )

where aiias was negiected to sirnplify the calculation. For lü = 201, the spectrum of the box kernel can be approxi-rnated by a pyramid consisting of 8 levels, as shown in Figure 3. Mapping of the l inear enhancement shown in Figure1 is now easiiy done by applying the factor CFo - SF." ' CF," to the highpass image and CF; : CF,., 0 < i < 7 ,to ttre other subband images. The lowpass subband image remains unchanged. Unsharp masking algorithms withdensitl, or contrast dependent gain factors are mapped similarly.

3. SUBBAND PROCESSING

Clinical acceptance depends decisively on ar appropriate design of the subband enhancement and careful tuningof its paramelers. In this respect backward compatibility as achieved by the above mapping, which allows re-useof introduced pararneters, is an important issue. Moreover we irnplemented and evaluated a set of new features,described in the following sections.

3.1. Structure Boost

This feature is provided by amplitude-dependent contrast amplification appiied to high- and bandpass subbandimages. The contrast amplification function has been designed according to the foliowing requirements:

r For large contrasts the function should approach the constant value C4 according to (2).

r For contrast values close to zero tire function should approach CF; + Gt, Gt > 0. G, defi.nes the additionalgain for weak contrasts within subband image i. For Gr :0, 0 1i 1n - 2 Structure Boost is disabled andour algorithm behaves iike linear unsharp-masking in Figure 1.

A suitable amplification function is given by

The additional enhancement is restricted to weak contrasts between zero and the transition amplitude c6. The gain

Gt allows a smooth transition towards the linear unsharp-masking processing. The exponent p defines how fast theamplification factor decreases towards CF;. Modification of the exponent does not influence the amplification ofcontrast values close to zero where also most of the noise can be found. Therefore, the exponent parameter can be

used to control the amount of additional weak structure enhancement whiie being to some extent robust with respectto noise amplification. The dependence of the contrast amplification on gain G is shorvn in Figure 4. The clinicalevaluation showed that modification of (" and p is of practical use, whereas for the transition amplitude co a constantvalue can be chosen, u'hich depends only on the number of quantization levels (typical ly between 10 and f6 bit) ofthe radiographs.

The effect of Structure Boost is demonstrated in f igure 5, showing a skul l i rnage processed with standard processingand multiscaie processirrg. Tire image was obtained in one of our clinical studies and q,as taken to examine the stateof the patient after a fal l . Tire radiograph exhibits an occipital fracture of the calotte. According to the Radiologists'commeDts the rnult iscale processed irnage provides a superior representation of small structures rendering an almost3-d l ike impression of the skui l . The visibi l i ty of the fracture l ine has been clearly improved. Thls holds also for thestructure of vesseis on the calotre. Enhancernent of the highpass subband irnage was reduced compared to the othersubband images in order to avoid noisJ' or spider net l ike appearance of the bone structure of the calotte. The lattermight feign a pathology.

(4 )CF,(c:\ = [ G, (r - $)" ' + cr, for lc l < ca

" [ C 4 f o r l c l > c 6

2.5

G - 1 n

- . = * - = - - - , . : - . ,

*I'ii ro

local inDut contrast

Figure 4. Contrast amplification function according to (4) for co = 30, P - L.5, C Fi = 1.0, and various vaiues of G

Figure 5. Scul l processed with standard unsharp-masking algori thm (left) and mult iscale algori thm (r ight).

3 .2 . No ise Conta inment

The Noise Problern. As X-raf images are usuallv subjected to a logarithmic gain curve during detection, noise

power is strongest in areas wilh lorv incident X-ray dose, r,r , l t ich are drsplayed as trr ight areas. Addit ional lv ' noise

is rnainly high äequent. Hence. rroise affects the radiograph predominairt ly in high-frequency bands of low optical

clensity. Subjer:t iveh', noise is rnost strongly visible in areas of weak or no texture-caused intensit l ' f luctuatrons

("activi ty"). The 3-r l representatiorr in Figure 6 combines this perceptual effect * ' i th the physical clependence of

t . 3

oaoco6

. o=E-E(g

0 . 5

40l 51 O

notse, inci icat ing the region where noise is most disturbing. Basical ly, enhancernent of (weak) detai ls also ampli f ies

noise' Based on the above considerations about noise sensitive regions, our noise containment strategy seeks to keepnoise within acceptable limits without losing the benefits of weak structure enhancement.

Figure 6' Quali tat ive piot of subjective noise visibi l i ty dependent on signai act ivi ty takrng into account physicalpresence of noise. The spatial frequency is subdivided into n octaves according to the subband images So, S, ,q - t - .u n - l r r 5 ( 1 2 - l l .

Realization of Noise Containment. Measures for local densitv and local activity allow to locate each subbandirnage pixel within the diagram of figure 6. Pixel values mainly affected by noise are Iocated close to the edgegiven bv lowest density, lowest activity and highest spatial frequencies (grey tretrahed.reon). The effecti'eness ofour noise control algorithm is focussed on this most noise sensitive region by integrating a density-, activity-, andscale-dependent control into the enhancement. Only the amount of enhancement is conträlled and no informatiorr rstaken away' from the original image. w'e hence refer to it as Noise containment.

A convenient way of implementation is to first apply the enhancement function to the subband pixel and thenweight the difference value between enlianced and original pixel by an attenuation factor b(x,g) thatäepends on thcsubband index i (= spatial frequency), local image density and local act ivi ty. The attenuation factor b(r,ü is givenby t lre product of densitv and activi ty dependent factors be(t,u) and öp(r,y) such that i t is smaller than I in thenoise sensitive region and equal to 1 elsewhere. A block diagram of the algoriit - i, given in Figure 7. The enhancedsubband st,f , ,r tv.r l , är ld the original subband s;,6, in är€ *. ight"d and added accoräing to

Sr ,n , . , ,h . ( r , a ) : b ( r , A) . S ; . f uuo " , ,h . ( r , A) + O - b ( r , AD . S i ,o . rg ( r , ! l )

providing a rroise robust pixel value Sr.,r, "rrn.(r, !) .The described method is easy to implement and the concept works regar<i less of the type ( l inear, nonl inear, . . .)oI er)nancemenl Guttlance of Noise Containment by imaging physics as weli u, p.r."pirul aspects allows a wellfocussed restr ict ion to the most noise sensit ive regions which normally consti tute a relat ively small part of the image(f igure 6) On the rest o{ the irnage the ful l potential of contrast elharrcement can be exploited.

. Tlre local density rreasure IVp can easi ly be retr ieved from the expanded and enhanced image of t i re next lowerlevel of the pyramid Iett .s(rqr),etp. Atr example of bp versus Ä:Ip is given in Figure 7. The density-clepenclentweighting factor öp is one in dark areas (high optical density), corresponding to ful l contrast enhancement. Start ingat a cri t ical densitY value Mp,", i t t lecreases over the interval A,{/p towards a value bD.^in corresponding to aresidual enliancement- This transition, however, clepends on tire subbancl: whereas in higi-fiequency subba.ds forl ^ -tuu' ( lel lsl t les a recluct ' ion of eni iancement might be necessarr,, i t rvi l l not be neecied in io*er-frequency subbands,, . . I - : , rwtrr( ' I l l )roYlcle better signal to noise rat ios. For b1 versus,41.4 sirr i i lar considerations hold. Activi ty measures ca'l . ^' r t talcul i t ted ustttg' ior instattce, t i re standard deviat iorr or local one-dirnensiolal histoqram-based measures l ike

( 5 )

D D,- in

0.0

^ M o

0.0 M h ^ 1 . 5 3 . 0

l@al densitY

Figure 7. Diagram Q of noise containment algorithm

Figure g. Dorsal spine processecl u,ith multiscale algorithm without (left) and with Noise Containment (right)'

information (negative entropy) or energy. The lateral radiograph of a dorsal spine demonstrates the effect of Noise

Containment (I'igure g). A good representation of all regions of the lateral spine must cope with the large density

variations. I' the ,rppe. rrÄndary area the shoulder-blades cause disturbing superpositions, while in the lower

boundary area the large densit l ' t ransit ion between lung and abdominal cavity must be handled. In both boundary

areas. high noise le'eis degrade the image impression. Ünacceptabie noise boosting on the f inest scales, part icularlv

in the bright regions of the upper ancl Iower parts of this radiograph, coulcl successful ly be prevented througli Noise

Clontainrnent. Expert crornmerrts included:

e Structure Boost enabies better exploitat ion of the avai lable r iynanric range of the f i lm together with excel lent

representatiol of rveakly' contrasting detai ls such as the posterior edges of the bodies of the vertebra'

r Noise Containment provides distinct quality improvement in the lorver ancl upper boundary area of the dorsal

sprne.

o Noise Containment rloes not lead to anY loss of detail inforrnation

3.3. Density-Controlled Contrast Enhancement

Requirements on contrast enhancement may depend on the region of the image because of e.g. the presence of noiseor the risk of feigned pathologies due to over-enhancernent. The latter is the case e.g. for chest imagcs. Here anamount of c<tntrast etrliancentent that could weli be applied to the region of the mediastine couid be unacceptablc:for the region of the iung.

Concerning the lung area, chest image processing should provide good contr ixt ofthe pulrnonary structures, rendervisible weakly contrasting pathological structures, e.g. lung nodules, and visual ize continuity of bone stmr:ture. Theprocessing should avoid to over-emphasize the pulmonary vascular system which could feign pneumonia. The retro-cardiac space should become more transparent e.g. to improve visibi l i ty of superposed lung structure. Another go:r lis a good transparenclr of the mediastine to al low dist inct ion of bone structure and vesseis, trachea and bronchi,posit ioning of catheter, wire, art i f ic ial cardiac valves, spinal column and the shadow of the heart. The processirrgsliould avoid noise amplification in abdomen, mediastine and retro-cardiac space.

A cl inical study at the Medical University of Hannover (see also section 3.5) confirmed that density-, and scale-dependent control of enhancement similar tci the density dependent part of Noise Containment is well suited to copewith these processing goals. This is because different anatomical detaiis of chest images are separable by the criteriaof detail size and local irnage density. It has to be noted that a prerequisite for the separation of anatomical reg;ionsof the chest image by iocal density is a ranging algorithm that automatically adapts the density curve to the actualexposure of tlie detector such that characteristic regions are always mapped to the same local density. This is stateof the art in current products. l3

region, detai l local density scales required contrast enhancement

abdomen, noise < 0 .5 0 , 1 no enhancementabdomen. relevant details < 0 .5 > 2 moderate to strons enhancementmediastine & retro-cardiac. noise < 1 . 0 0, 1 no enhancementrnediastine & retro-cardiac. relevant detai ls < 1 . 0 > 2 moderate to strong enhancementung, rib contours - t . 7 0 , 1 moderate enhancementung, pulmonarl' vascular system - 7 . 7 t a low enhancementung. la rge weak lv cont ras t ing r rodu les - 1 . 7 > 4 moderate to strons enhancement

Table 1. IdentificrLtion of anatornical regions within chest images by scale and local density, and the requirementson contrast enirancement, respectively. The statements hold for a spatial resolut ion of 2.5 lp/mm.

The identification of anatomical details of chest images is listed in Table 1. The requirements on contrastenhancemetrt result from the processing goais mentioned above. Also, noise sensitive regions can be identified,which leads to a cyuite satisfactorf implementation of Noise Containment as a spin-off feature of Density-ControlledContrast Enhancement. Density-r:ontrolled contrast enhancement is realized by integrating a density- and scale-dependent control into the enhancenient u,'hich influences the amount of enhancement. This aigorithm correspondsto Noise Containment (f igure 7) without the activi ty-dependent control path. The subband pixel is f i rst enhancedaccording to the contrast amplification function (4). Then the difference value between enhanced and original pixeiis weighted bv an attenuation factor 0 that depends on the subband index i (= spatiai frequency) and local intagedensitl'. This n'eightecl difference value is finally added to tlie original subband pixel. The attenuation factor ü isselected such that i t is smaller than 1 in regions of reduced enhancernent and equal to 1 in regions of maximumenhancement. This croncept works regardless of the type ( l inear, nonl inear, . .) of enhancement. Results of thedensitv-control lecl nruit iscale processing were compared with a standard chest processing.r3 The weighting factorsf<-rr nlr l t iscale corrtrast enirancenient were determined at t l ie Meclical University of Hannover (f igure 9). Contrastwas gerieral l l ' . judged to be better in the mult iscale processecl irnage, especial iy in the brigirter regions. Due to theselectiotr oi the densitv curves. the processing did not disturb the balanced image impression, and did not feiglnon-exist ing pathologies. A snral l dlarvback rvas a sl ight increase in noise that could not be avoided.

-t4/' /

//

/

../ suooanao +-'

subband 1 -''*--subbands 2, 3

subband 4 "subband 5 - - - . - -

local density

Figure 9. \ ! 'eighting factor ö for density-control led mult iscale chest processing.

3.4. Equal izat ion

We <lefine Equo,Iization as scale-dependent variation of the contrast amplification function. Contrary to density-

<rontrolled contrast enhancement, which can aiso vary from scale to scale but only affects certain regions of the image.

Equulizat'ion influences the overall impression of the image. Apart from the scale-dependent factors CF;, we have

for practical reasons restr icted scale dependence to the gain factor G; in (a). While easv to handle, this produces

r:linically vaiuable results without causing visible artifacts. Preference of higher spatial frequencies, preference of

lower spatial frequencies and equal treatment of all scales are three typical Equalizer settings for the gain that we

determined during our ci ini<:al studies.

The preference of the subband image corresponding to the smallest details provides a sharper image impression.

This rvas the preferred parameterization for images where the diagnostic interest was exclusively in small details iikel l^.p rpytrrrc Strnnger enhancement of lower scales would only produce disturbing density variat ions result ing from

sriperposed soft tissue. In such images, the additional boosting of noise could be accepted because noise is masked

bv high frequent texture infbrmation.

The eclual treatnrent of all scales was preferred for images where the diagnostic interest was equally spread from

r,r:rv small to large details. Superposition effects should not be reduced but structures of different sizes should become

easier to distinguish. Boclstinpi of the noise level could irritate in some regions of the image. In this case the noise

corrrpensation mechanism proved to be a suitable countermeasure. Typical examples are lateral spine images (see

figure 10). Equal treatment of all scales is appropriate because diagnostic interest targets small details, such as the

posterior edges of a vertebra, as rvel l as larger detai ls l ike a body of a vertebra. Unacceptable noise boosting on the

finest sr:ales, part icularly in the bright regions of the upper and lower parts of this radiograph, could successful ly be

I)revented thr<iugh the Noise Containlnent options.

If the imlge contains few snial l sized detai ls of diagnostic interest, preferred enhancernent of large sized detai ls

is a good lneans to prevent too mucli noise ampli f icat ion (see Figure 6). Cl inicians found this to give a three dimen-

siorral i rnpression to the images. Pelvis images, for instance, were found to iclok similar to conventional (analogue)

raci iographs wit irout loosing the advantage of better exploitat ion of the avai lable dynamic range. Erthancement of

weakll' r:ontrasting medium sized cletaiis is clesirecl, while the same enhancernent applied to the finest scales was said

to gelerate an "unnatural" irnagc impression. This led to the contrast ampli f icat ion functions in Figure 10. The

srnooth tr2nsit ion of corrtrast gain values over scale was judged to give a balanced image impression.

3.5. Cl inical Studies

f irc rlultiscale algorithm was evaluated orr a storage phosphor radiography system in the radiology departments of

F\rlfla N4unicipal Hospital, and crn a seleniurn-based chest-radiographv system at the Medical Urriversity of Hannover.

A r: lr l ical prototype incorltorat ing the cornpiete functional i ty of a diagnostic workstat ion was connected to the

141 4

o6 0 .6o--o

=EO N '

6 w . +

3

o.2a2 .5t _ 30 .5

--JD,

ampl . tac t . 'dorsa l sPrne

3 .53

2 .52

1 . 51

0

Figure 10. Contrast arnplification function for dorsai spine (left) and pelvis (right) as dependent on contra^st

amplitude alrd scale index.

radiography systems, al lorving continuous operation in the diagnostic routine. A comprehensive archiving solut ion

allowed consistent storage of inrages, processing parameters and radiologists' comments. For printing, images were fed

back to the hospital workstation. Evaluation of the images was exclusiveiy made on film printouts. On the prototype

Figure 11. Design of subband enhancement'

the ciescribed algori thm rvas reai ized as fol lows (see Figure 11). Prior to mult iscaie contrast enhancement, a l i im-

like look-up tabie was applieci to the radiographs. Then the images were subjected to a pyramidal decornposition

accordinpi to the compatibility criteria clescribed in 2. For linear contrast enhancement and additional Structure

Boost al isubbancl images S;, 0 < i <, (rt - 1) were subjected to a remapping operation. The corresponding look-up

table transforrns the contrast values according to:

( t , , , " - oppe , t ( r , l ) = C F ; ( c i ( " , 3 / ) ) ' c i ( t , 9 )

CF;(ci(x,E)) obeys (a). Noise Containment is control leci by the local density Ineasure only, so that this feature

could be real ized via density-control ied contrast enhancernent. (see section 3.2). The density measure can easi iy be

retrieved frorn the expancled reconstructecl image of the next lower level. The clinical studies were divided into two

phases: one for pa.a-ete, optimizatiol uncler a radiologist's guiclance, and one to collect feedback on the processing

resuits from a larger number of radiologists.

During the first phase the irnage iurpression was optimized by exploiting the modification options provided bv

our algori,hm. StaUilitv of the obtained paramete., ** verifiecl by throughput of a iarger number of radiographs'

(6 )

local input contrasl

1415

We covered the major part of borre diagnostics as well :r^s chest ancl abdorninal regions. rhis way we collecte4 aconsiderable number of routine radiographs, optimized parameter sets, and expert cornrnents r11 ca-se l istories, irnagequali ty ' and processing performance. The comments were arrangecl in :rn HTML-based clgc'mentati .r i , with thclnlages serving as active l inks. Cl icking into the images autornatical ly loads them into the prototype a'ci act ivatesthe corresponding parameter set, thus permitt ing recreation of the cl inical tr ials.

In the case of the study on the storage phosphor system, twenty images processed by both standarcl processrrrgand mult iscale processing, using the optimized parameters, were presented side-by-side to 1.1 radiologists, with<tutidentifying to thern rvhich ones were multiscale-processed, and which ones were standard-processed. Tire comparativeevaluations were ba^sed on the criteria in Figure 12, defined by the radiologists well before the session starteci. Ttrr:

Figure 12. Opinion scores for dorsal spine, averaged over all radiologists (left). Scores averaged over all radiologistsand al l evaluated examination types (r ight).

right part of diagram 12 summarizes the results, showing grand averages over all images and all readers (broadbar). The averages are based on numerical values ranging from -3 to *3. Addit ional ly,the standard deviat ion ofeach grand average ','alue is given (narrow bar). Muitiscale processing outperforms the standard algorithrn almosteverywhere, without affect ing the subjective cri terion of "harmonic image impression". The increased e.hancementpotential did evidently not result in a related boost of noise. Scores for the last three criteria (critical areas, principaland other f indings) shoq'that there is indeed good reason to bel ieve that the new processing eäses t5e cl iagnosticprocess' The study aiso revealed that the achievable quality enhancement depends on the imagecl anatomical iegion.For some re6iions like the hand, results from standard processing are already close to the optrmum, since aimost nohidden detai ls could be rnade better visible. For other regions such as ciorsai spine, mult iscale processing signif icantlyimproved image quality (see left part of diagram 12), rnuch better in fact than the averaged scores sho*. Hence, aslong as the focus is on harmonization and firre detail enhancement in not too noisy conditions, the described referencealgorithm is already a po'"r''erful processing tool. Due to backward compatibility, multiscale processing should alwaysbe at least as good as this reference algorithm.

In the case of the study on the seleniun-basecl chest system, images of an anthropomorphic chest phantoni thatw'ere randomlv superimposed with simulated lesions were processed b1' chest standard processingl3 anci mult iscaleprocessing. Detection of lesions was assessed by f ive radiologists and evaluateri by Jackknife ROC-methodology. Thestudy showed t l tat rnult iscale enl iancement rvas signif icantly superior for t i re detection of micro-rrodules in the lungarea of adipose patients (A' 0.82 vs. 0.76). For other lesiorr types investigated, dif ferences were not signif icant,though there rvas a general trend towards superior detectiorr with mult iscale enhancement. i4

4. DISCUSSIONW'e developed a nerv nonlinear multiscale algorithm that canprocessing. Backward compatibi l i ty enables re-use of processingand eases acceptance of the new algori thm. Part icuiar attention

be derived from unsharp ma^sking-based standardkrrow-how acquired on current stanclard processingwas paid to noise robustness.

A clirrical study confirms that a suitable design of the subband processing vielcls a significantly improvecl detailvisibility for a broad spectrurn of routine radiographs rvithout unacceptable boosting oi loise and distortion of anatura l image impress ion .

Detectors of the next generation as e.g. sol id state cietectors wil l br ing further improvements rvith respect toversati l i ty, sigrral to t toise rat io, dynamic range, and spatial resolut ion.l5,16'1 We e*peci that the advantages of thenew processing, especial ly the specif ic enhancement of weakly contrasting structures of any size, wi l l be indiJpensableto exploit the ful l potential of these detectors.

ACKNOWLEDGMENTS

We thank the director, Prof. NI. Galanski, and the staff of the Institute of Diagnostic Radiology of the MedicalUniversity of Hannover for their support. We thank the director, Prof. J.-P. Hais, ancl the staff of the racliologvdepartment of Fulda Municipal Hospital for their support. We are especiall l ' grateful to Dr. E. Dencker ancl Dr. E.\4üller for their helpful contributions and their patience during many extencled sessions. \l'e furthermore tharrk ourcolleagues at Phil ips Medical Svstems, Hamburg, for fruitful discussions, technical l ielp and for provicling applicationalbackground.

REFERENCES

T. Aach, U. Schiebel, and G. Spekowius, "Digital image acquisit ion and processing in rnedical x-rav lmaging,',Journal of Electronic Imaginq (Special Section on Biomed,ical Image Representation) g(I),199g (in p..r.;.

-

N{. Ishida, Computed Radiogruphy (Tateno, I inuma, Takano, eds.), ch. IrnapJe Processing, pp. 25_�30. SpringerVerlag, Tokyo, 1987.

3. I. N'Iaack and U. Neitzel, "Optimised irnage processing for routine digital radiography,,'in Proceed'ings ComputerAssisted Radiology, U. Lemke, ed., pp. 108-114, Berlin, Springer Verlag, 1gg1.

4. A. R. Cowen, A. Giles, A. G. Da'ies, and A. Workman. ,,An image processingin SPIE Vol. 1898, Irnaqe Processing,pp.833-843, 1993.

algorit i rm for PPCR imaging.

5. T. L. . I i , \ '1. K. Sundareshan, and H. Roehrig, "Adaptive iniage contrast enhancement basecl on human visual

l 1

properties," IEEE Transo.ctions on Medical Imaging 13(4), pp. 573-5g6, 1994.A. G. Davies, A' R. Cowen, G. J. S. Parkin, and R. F. Bur, "Optimising the processing anrl presentation ofPPCR images," in SPIE VoI. 2712, Image Processing, pp. l8g-195, 1996.A. N. Akansu and R.- A. Haddad, Multiresolution Signal Decomposition^ Acaclemic Press, Boston. 1992.P. J. Burt and E. H. Aclelson, "The Laplacian pyramid as a compact image r:od,e," IEEE Transactions ortCommu.nications 31(4) , pp. 532-540, 1g83.R. F. Bessler and .I. H. Arbeiter. "Contrast enhancement using Burt pyramid processing," in SID Internr1iottulSym\tosium,, Digest ol Techntcal Papers, pp.352-353, (San Diego) , Mav 6-g 1gg6.P. \ ltvisteke arrd E. Schoeters. "Multiscale image contrast amplif ication (NIUSICAl",v'1,, ' i1 SpIE Vol I j67.Irnage Proce.ssing, pp. 551 560, 1994.S. Ranganath and H. Blume. "Hierarchical image clecornposition anci f i l tering usirrg the S-transform.', in ^gp1EIro l . 914. pp. 799-814, 1988.S. Ranganath, "Image filtering using multiresolution representations," IEEE Trunsactions on pattern Analusi.sand Mach,ine Intell igence 13(5). pp. 420 '140, 1991.H Blume. "Digital processing of THORA\IISION images: The Algorithms," prorluct clescription, phil ips lvledic3lSysterns, 1996.E. Dencker, NI. Staii l , S. Dippel, C. M. Schaefer-Prokop, S. Baus, S. Goehde, D. l loegemann, A. Leppert. andX'I. Galanski, "Noniinear rnultiscale processingin selenium-basecl chest ra<iiograph1,." in ECR, Book of Abstract,s.1999 .N Jung, P- L. Alving, F. I lusse, N. Conrads, H. Ir,1. Ivleulenbrugge, W. R.ütterr. U. Schiebel, M. Weibrer:1r, iurciH. Wieczorek, "Dynatnit x-ra,\ ' irnaging system based on an amorpirous sil icon thin-fi lm array," in SpIE Vol.; t3 , t6.1998.IJ. Sclriebel. N. Conracls. N. Jung, M. \! 'eibrecht. H. \Vieczorek. T. Zaerge),. \,I..J. powell, I. D. Frencli, anclC. Cllasse. "Fluoroscopir: imaging with amorphous sil icorr thin-fi lni arrays," in SplE Vol. 2163. pp. 129-140.199,1.

l 5

7 .8 .

Lt)

t 2

t . l

1.1.

16